MLLog.dev is a machine learning log: in-depth breakdowns of the latest ML and AI papers, hands-on tutorials, experiments and projects.https://mllog.dev/en-USMaciej Myszkiewiczhttps://mllog.dev/en/posts/orca-world-foundation-model-next-state-prediction/https://mllog.dev/en/posts/orca-world-foundation-model-next-state-prediction/Orca (BAAI) replaces next-token, next-frame, and next-action prediction with a single Next-State-Prediction objective. A frozen 4B backbone pre-trained on 12.5K hours of video — with zero action labels — feeds three lightweight readouts, and the action readout, trained on just 200 trajectories per task, beats π0.5 on OOD robot manipulation (32.4 vs 29.4).Sun, 05 Jul 2026 07:00:00 GMTworld-modelsnext-state-predictionrepresentation-learningroboticsmultimodalhttps://mllog.dev/en/posts/swe-explore-benchmarking-coding-agents-repository-exploration/https://mllog.dev/en/posts/swe-explore-benchmarking-coding-agents-repository-exploration/SWE-Explore isolates repository exploration from patch generation, revealing that coding agents find the right files ~65% of the time but recall only ~15-19% of the lines that actually matter — and that context efficiency predicts downstream resolve rate with Pearson r = 0.950.Tue, 09 Jun 2026 07:00:00 GMTcoding-agentsbenchmarkscode-localizationsoftware-engineeringrepository-explorationhttps://mllog.dev/en/posts/skillopt-text-space-optimizer-agent-skills/https://mllog.dev/en/posts/skillopt-text-space-optimizer-agent-skills/SkillOpt applies deep-learning-style optimization - bounded edit budgets, validation gating, rejected-edit memory - to natural-language skill documents, improving frozen LLMs by up to +39 points across 52/52 evaluated cells without changing a single model weight.Fri, 29 May 2026 07:00:00 GMTagentsprompt-optimizationllmskill-learningtext-space-traininghttps://mllog.dev/en/posts/molmoact2-action-reasoning-real-world-deployment/https://mllog.dev/en/posts/molmoact2-action-reasoning-real-world-deployment/MolmoAct2 is a fully open vision-language-action model that outperforms π0.5 and matches Gemini Robotics ER, achieving 97.2% on LIBERO and 87.1% real-world success via per-layer KV-cache conditioning and adaptive depth reasoning.Sun, 10 May 2026 11:00:00 GMTroboticsVLAembodied-reasoningflow-matchingopen-sourcehttps://mllog.dev/en/posts/recursive-multi-agent-systems/https://mllog.dev/en/posts/recursive-multi-agent-systems/RecursiveMAS treats an entire multi-agent system as a single recursive computation in latent space, adding only 0.31% trainable parameters while achieving +8.3% accuracy, 2.4x speedup, and 75.6% token reduction over text-based multi-agent baselines.Sat, 02 May 2026 00:00:00 GMTAImulti-agent-systemsrecursive-computationlatent-spacellm-collaborationhttps://mllog.dev/en/posts/2026-04-26-tstars-tryon-virtual-try-on-mmdit/https://mllog.dev/en/posts/2026-04-26-tstars-tryon-virtual-try-on-mmdit/How a unified 5B MMDiT trained with multi-reward RL and step distillation reframes virtual try-on as multi-image editing — and runs in under 4 seconds in production.Sun, 26 Apr 2026 00:00:00 GMTAIdiffusionvirtual-try-onMMDiTimage-editingdeploymenthttps://mllog.dev/en/posts/2026-04-15-clawgui-unified-framework-gui-agents/https://mllog.dev/en/posts/2026-04-15-clawgui-unified-framework-gui-agents/ClawGUI unifies online RL training, reproducible evaluation, and real-device deployment of GUI agents into one open-source pipeline — and shows a 2B model trained inside it can beat 72B untrained baselines on MobileWorld.Wed, 15 Apr 2026 09:00:00 GMTgui-agentsreinforcement-learningbenchmarkingdeploymentprocess-reward-modelhttps://mllog.dev/en/posts/2026-04-12-skillclaw-collective-skill-evolution-llm-agents/https://mllog.dev/en/posts/2026-04-12-skillclaw-collective-skill-evolution-llm-agents/SkillClaw is a framework for collective skill evolution in multi-user LLM agent ecosystems. Instead of static skill libraries, the system automatically learns from interactions across users and propagates improvements to everyone.Sun, 12 Apr 2026 00:00:00 GMTllm-agentsskill-evolutionmulti-userself-improvementagent-systemshttps://mllog.dev/en/posts/taps-task-aware-speculative-decoding/https://mllog.dev/en/posts/taps-task-aware-speculative-decoding/TAPS shows draft training data drives speculative decoding: merged-tree verification of math and chat specialists hits 5.11 acceptance length, weight averaging 2.59.Sat, 28 Mar 2026 00:00:00 GMTAINLPLLMSpeculative DecodingInference OptimizationEAGLEHASShttps://mllog.dev/en/posts/demystifying-video-reasoning-chain-of-steps/https://mllog.dev/en/posts/demystifying-video-reasoning-chain-of-steps/Chain-of-Steps shows video diffusion models reason across denoising steps, not frames, exploring and self-correcting before converging on a single answer.Tue, 17 Mar 2026 00:00:00 GMTAIComputer VisionVideo GenerationDiffusion ModelsReasoningDiffusion Transformerhttps://mllog.dev/en/posts/seoul-world-model-city-scale-video-generation/https://mllog.dev/en/posts/seoul-world-model-city-scale-video-generation/Seoul World Model generates photorealistic city video from street-view photos, cutting FID on unseen Busan to 28.4 from 62.1 via cross-temporal reference pairing.Mon, 16 Mar 2026 00:00:00 GMTAIComputer VisionVideo GenerationWorld ModelsAutonomous DrivingDiffusion Transformerhttps://mllog.dev/en/posts/constory-consistency-bugs-long-stories-llm/https://mllog.dev/en/posts/constory-consistency-bugs-long-stories-llm/ConStory-Bench measures consistency in 10,000-word stories: GPT-5-Reasoning leads with 0.113 errors per 10k words, and its checker finds 3.2x more bugs than experts.Mon, 09 Mar 2026 00:00:00 GMTAILLMNLPBenchmarkText GenerationNarrative ConsistencyEvaluationhttps://mllog.dev/en/posts/utonia-one-encoder-all-point-clouds/https://mllog.dev/en/posts/utonia-one-encoder-all-point-clouds/Utonia trains one 137M encoder for indoor, LiDAR, satellite, CAD and video point clouds. Stripped of color it scores 77.0% mIoU on ScanNet vs Concerto's 36.8%.Sat, 07 Mar 2026 00:00:00 GMTAI3DPoint CloudsSelf-Supervised LearningRoboticsAutonomous VehiclesVision Transformerhttps://mllog.dev/en/posts/sage-reasoning-model-knows-when-to-stop-thinking/https://mllog.dev/en/posts/sage-reasoning-model-knows-when-to-stop-thinking/SAGE scores reasoning paths by cumulative log-probability instead of next-token probability, letting models stop thinking 40-50% earlier at equal or better accuracy.Mon, 23 Feb 2026 00:00:00 GMTAILLMReasoningReinforcement LearningOptimizationChain-of-ThoughtEfficiencyhttps://mllog.dev/en/posts/gpt-discovers-physics-gluon-amplitudes/https://mllog.dev/en/posts/gpt-discovers-physics-gluon-amplitudes/GPT-5.2 Pro proposed the formula showing single-minus gluon amplitudes are nonzero in Klein space, valued only in -1, 0 or 1; it passed five consistency tests.Sun, 15 Feb 2026 00:00:00 GMTAIGPTPhysicsOpenAIScientific DiscoveryGluonsField Theoryhttps://mllog.dev/en/posts/opus-efficient-data-selection-llm-pretraining/https://mllog.dev/en/posts/opus-efficient-data-selection-llm-pretraining/OPUS scores every sample by alignment minus redundancy in optimizer update space; on Qwen3-8B it beat full 3B-token training using 0.5B, at 4.7% extra compute.Fri, 13 Feb 2026 00:00:00 GMTAILLMMachine LearningDeep LearningData SelectionPretrainingOptimizationhttps://mllog.dev/en/posts/green-vla-staged-vision-language-action-generalist-robots/https://mllog.dev/en/posts/green-vla-staged-vision-language-action-generalist-robots/Green-VLA trains one policy for humanoids, mobile manipulators and fixed arms across five stages and 3,000 hours of demos, hitting state of the art on CALVIN ABC-D.Sun, 08 Feb 2026 00:00:00 GMTAIRoboticsVLAMachine LearningDeep LearningReinforcement LearningVision-Language Modelshttps://mllog.dev/en/posts/grokking-provable-ridge-regression/https://mllog.dev/en/posts/grokking-provable-ridge-regression/Grokking happens even in ridge regression: fast memorization, a long plateau of orthogonal noise, then late generalization with a delay that scales as 1/λ.Tue, 27 Jan 2026 00:00:00 GMTMachine LearningGeneralizationGrokkingRidge RegressionRegularizationTheoryhttps://mllog.dev/en/posts/tensor-networks-neuro-symbolic-ai/https://mllog.dev/en/posts/tensor-networks-neuro-symbolic-ai/Tensor networks give neural and symbolic AI one language: logic becomes sparse tensors, neural nets low-rank ones, and inference in both is a contraction.Fri, 23 Jan 2026 00:00:00 GMTMachine LearningNeuro-Symbolic AITensor NetworksLogicProbabilistic ModelsHybrid AIhttps://mllog.dev/en/posts/m2fmoe-extreme-adaptive-time-series-forecasting/https://mllog.dev/en/posts/m2fmoe-extreme-adaptive-time-series-forecasting/M2FMoE routes Fourier and wavelet experts by frequency band to forecast floods without extreme-event labels, cutting error 45.7% on Almaden and beating 13 baselines.Wed, 14 Jan 2026 00:00:00 GMTMachine LearningTime SeriesMixture of ExpertsForecastingFrequency AnalysisAAAI 2026https://mllog.dev/en/posts/ballast-contextual-bandits-raft-timeouts/https://mllog.dev/en/posts/ballast-contextual-bandits-raft-timeouts/BALLAST replaces Raft's random election timeouts with a LinUCB contextual bandit; a Rust simulator cuts cluster unavailability by 84%, and 90.8% under bursty WAN.Mon, 05 Jan 2026 00:00:00 GMTMachine LearningDistributed SystemsRaftContextual BanditsLinUCBhttps://mllog.dev/en/posts/ai-co-scientist-rubric-rewards/https://mllog.dev/en/posts/ai-co-scientist-rubric-rewards/Rubrics mined from papers turn a frozen grader into an RL reward; experts preferred the generated research plans 70% over baseline and 52% over published ones.Tue, 30 Dec 2025 00:00:00 GMTMachine LearningLLMReinforcement LearningAI in ScienceResearchhttps://mllog.dev/en/posts/hydra-dynamic-rank-adaptation-mobile-vlm/https://mllog.dev/en/posts/hydra-dynamic-rank-adaptation-mobile-vlm/HyDRA allocates LoRA rank per layer and inside layers automatically, buying a 4.7% gain on mobile vision-language models with no extra trainable parameters.Sat, 27 Dec 2025 00:00:00 GMTMachine LearningLoRAFine-tuningVLMMobile AIOptimizationhttps://mllog.dev/en/posts/comp-llm-composable-inference-framework-analysis-2025/https://mllog.dev/en/posts/comp-llm-composable-inference-framework-analysis-2025/Comp-LLM routes a query as a DAG of sub-queries to parallel experts: about 35B scores 0.83 F1 to Llama-2 70B's 0.85, and 1.1-1.7x faster than sequential agents.Tue, 09 Dec 2025 00:00:00 GMTAILLMMachine LearningComp-LLMArchitectureResearchhttps://mllog.dev/en/posts/nvidia-nemotron-parse-v1-1-analiza-vlm-ocr-deep-dive/https://mllog.dev/en/posts/nvidia-nemotron-parse-v1-1-analiza-vlm-ocr-deep-dive/Nemotron Parse v1.1 pairs a C-RADIO ViT-H encoder with an mBART decoder in 885M parameters, squeezing 13,184 visual tokens to 3,201 and emitting Markdown or LaTeX.Wed, 26 Nov 2025 00:00:00 GMThttps://mllog.dev/en/posts/llm-cascades-cost-constrained-c3po/https://mllog.dev/en/posts/llm-cascades-cost-constrained-c3po/C3PO runs LLM cascades under a probabilistic cost constraint: stay in budget with 95% confidence, escalating to a bigger model only when the cheap one falls short.Fri, 14 Nov 2025 00:00:00 GMTAILLMMachine LearningOptimizationC3POhttps://mllog.dev/en/posts/accurate-satellite-rain-forecasting/https://mllog.dev/en/posts/accurate-satellite-rain-forecasting/TUPANN nowcasts rainfall from satellite data alone, adding a differentiable advection operator and lead-time-conditioned MaxViT for forecasts 10-180 minutes ahead.Mon, 10 Nov 2025 00:00:00 GMTAIweathermachine learningphysicssatellitehttps://mllog.dev/en/posts/pinfm-foundation-model/https://mllog.dev/en/posts/pinfm-foundation-model/HYSTL pairs a crime knowledge graph with a hypernetwork that generates GNN weights per crime type, letting one model forecast across cities with disjoint taxonomies.Thu, 06 Nov 2025 00:00:00 GMThttps://mllog.dev/en/posts/chord-smart-on-device-recommendations/https://mllog.dev/en/posts/chord-smart-on-device-recommendations/CHORD sends each phone a bit map from a cloud hypernetwork, quantizing a sequential recommender channel by channel so weak devices keep accuracy and battery.Mon, 06 Oct 2025 00:00:00 GMTAIrecommendationedge computingMLquantizationhttps://mllog.dev/en/posts/dialtree-rpo-red-teaming-multi-turn-attacks-llm/https://mllog.dev/en/posts/dialtree-rpo-red-teaming-multi-turn-attacks-llm/Friendly explanation of 'Tree‑based Dialogue Reinforced Policy Optimization for Red‑Teaming Attacks' — with examples and LaTeX formulas.Sat, 04 Oct 2025 00:00:00 GMTAIred-teamingLLMreinforcement-learningsecurityhttps://mllog.dev/en/posts/teaching-reasoning-models-to-explore-smarter/https://mllog.dev/en/posts/teaching-reasoning-models-to-explore-smarter/AttnRL uses a model's own attention to pick branching points in process-supervised RL, beating GRPO and TreeRL by 1.8 points on math while cutting training time 8%.Wed, 01 Oct 2025 00:00:00 GMTAIreinforcement learningLLMreasoningAttnRLhttps://mllog.dev/en/posts/how-to-detect-credit-card-fraud/https://mllog.dev/en/posts/how-to-detect-credit-card-fraud/Credit card fraud is 492 cases in 284,807 transactions; pairing undersampling with SMOTE beats either alone, and KNN and MLP gain most on the imbalanced test set.Sun, 21 Sep 2025 00:00:00 GMTAIfraud detectionmachine learningclass imbalanceMLPSMOTEhttps://mllog.dev/en/posts/quantum-trading-ai-quantum-computing/https://mllog.dev/en/posts/quantum-trading-ai-quantum-computing/A QLSTM forecaster plus a QA3C agent trade USD/TWD with just 244 parameters, returning 11.87% over 2020-2025 at 0.92% max drawdown versus classical A3C's 11.42%.Mon, 15 Sep 2025 00:00:00 GMTAIQuantum ComputingReinforcement LearningTradinghttps://mllog.dev/en/posts/reinforcement-learning-in-pinterest-ads/https://mllog.dev/en/posts/reinforcement-learning-in-pinterest-ads/DRL-PUT replaces Pinterest's hand-tuned ad ranking weights with a policy network picking utility weights per request; A/B tests show CTR up 9.7%, LC-CTR up 7.7%.Mon, 08 Sep 2025 00:00:00 GMTreinforcement learningrecommender systemsPinterestmachine learningadshttps://mllog.dev/en/posts/uczene_pinterest/https://mllog.dev/en/posts/uczene_pinterest/Pinterest's DRL-PUT learns the weights of its ad ranking utility with policy gradients instead of tuning them by hand, lifting CTR by 9.7% and LC-CTR by 7.7%.Mon, 08 Sep 2025 00:00:00 GMTreinforcement learningrekomendacjePinterestmachine learningadshttps://mllog.dev/en/posts/the-anatomy-of-ai-lies/https://mllog.dev/en/posts/the-anatomy-of-ai-lies/Can LLMs Lie? traces deception to layers 10-15 with logit lens, zero-ablation and steering vectors: models rehearse lies in dummy tokens; bigger models lie better.Fri, 05 Sep 2025 00:00:00 GMTAIArtificial IntelligenceLLMAI SafetyAI EthicsResearchInterpretabilityDeep Learninghttps://mllog.dev/en/posts/accelerate_neural_network/https://mllog.dev/en/posts/accelerate_neural_network/SLAC's SNL and Auto-SNL push neural network inference onto MPSoC FPGAs for LCLS-II, swapping weights without reconfiguration and matching or beating hls4ml latency.Mon, 01 Sep 2025 00:00:00 GMTAIFPGAHardware AccelerationMachine LearningPhysicsMPSoChttps://mllog.dev/en/posts/global-guarantees-of-robustness-probablistic-apporach/https://mllog.dev/en/posts/global-guarantees-of-robustness-probablistic-apporach/Rather than certify every point, Mu and Lim estimate the probability that a random input is non-robust, wrapping a Clopper-Pearson interval around a small sample.Wed, 27 Aug 2025 00:00:00 GMTArtificial IntelligenceMachine LearningAdversarial RobustnessGlobal GuaranteesAI Safetyhttps://mllog.dev/en/posts/intern-s1-ai-scientist/https://mllog.dev/en/posts/intern-s1-ai-scientist/Intern-S1 is an open-source MoE science foundation model with 28B activated parameters, trained on 5T tokens and a Mixture-of-Rewards spanning over 1000 tasks.Sat, 23 Aug 2025 00:00:00 GMTAIMachine LearningScienceMultimodal Modelshttps://mllog.dev/en/posts/seamless-flow-ai-factory/https://mllog.dev/en/posts/seamless-flow-ai-factory/Seamless Flow kills RL pipeline bubbles with a data plane and tag-driven GPU scheduling: up to 100% more throughput than VERL, and 23.0% to 45.8% on SWE-Bench.Mon, 18 Aug 2025 00:00:00 GMTAIReinforcement LearningLLMTech ExplainedGPU UtilizationSystem Designhttps://mllog.dev/en/posts/sok-data-minimization-in-machine-learning/https://mllog.dev/en/posts/sok-data-minimization-in-machine-learning/SoK: Data Minimization in Machine Learning maps the GDPR principle onto ML pipelines, separating horizontal from vertical techniques and pre-hoc from post-hoc steps.Fri, 15 Aug 2025 00:00:00 GMTmachine learningdata minimizationAIprivacyGDPRhttps://mllog.dev/en/posts/learning_machines/https://mllog.dev/en/posts/learning_machines/DyMoE counters catastrophic forgetting in incremental graph learning by adding a new expert per increment, with gating and a loss that makes older experts teach it.Thu, 14 Aug 2025 00:00:00 GMTArtificial IntelligenceMachine LearningGraphsSciencehttps://mllog.dev/en/posts/dynamic-fine-tuning-dft/https://mllog.dev/en/posts/dynamic-fine-tuning-dft/Dynamic Fine-Tuning rescales each token loss by the inverse of its probability, a one-line change that beats SFT on generalization and rivals offline RL methods.Mon, 11 Aug 2025 00:00:00 GMTAILLMMachine LearningSFTDFTReinforcement Learninghttps://mllog.dev/en/posts/deep-dive-text-to-sql-adaptive-method/https://mllog.dev/en/posts/deep-dive-text-to-sql-adaptive-method/Adaptive Text-to-SQL splits the work in two: a Selector retrieves conceptually similar question-schema-SQL triples, a Generator writes the query in-context.Mon, 11 Aug 2025 00:00:00 GMTArtificial IntelligenceNatural Language ProcessingSQLLLMText-to-SQLData AnalysisMachine Learninghttps://mllog.dev/en/posts/askdagger-interactive-imitation-learning/https://mllog.dev/en/posts/askdagger-interactive-imitation-learning/ASkDAgger lets a robot flag an uncertain plan before acting; adaptive gating and foresight replay cut teacher queries and raise success on CLIPort and a real arm.Fri, 08 Aug 2025 00:00:00 GMTAIMachine LearningRoboticsImitation LearningNLPhttps://mllog.dev/en/posts/capulse-teaching-machines-to-hear-the-rhythm-of-data/https://mllog.dev/en/posts/capulse-teaching-machines-to-hear-the-rhythm-of-data/CaPulse spots time-series anomalies by learning the causal rhythm of data: it fires when the cause-effect order breaks, not when a threshold is crossed.Thu, 07 Aug 2025 00:00:00 GMTAIMachine LearningScienceTechnologyTime Serieshttps://mllog.dev/en/posts/goedel-prover-v2-revolution-in-automated-theorem-proving/https://mllog.dev/en/posts/goedel-prover-v2-revolution-in-automated-theorem-proving/Goedel-Prover-V2 learns from scaffolded proof sketches and its own failed Lean attempts, outperforming AlphaProof and DeepSeek-Prover on miniF2F and PutnamBench.Wed, 06 Aug 2025 00:00:00 GMTAIArtificial IntelligenceMathematicsFormal ProvingLLMhttps://mllog.dev/en/posts/esoftmax-approximating-one-hot-vectors/https://mllog.dev/en/posts/esoftmax-approximating-one-hot-vectors/ε-Softmax swaps rigid one-hot targets for slightly softened ones, so a mislabeled example costs the model far less. A NeurIPS 2024 answer to noisy training labels.Tue, 05 Aug 2025 00:00:00 GMTAIMachine LearningResearchTechnologyhttps://mllog.dev/en/posts/quantitive_effective/https://mllog.dev/en/posts/quantitive_effective/Uncertainty from one deterministic model: kernel density estimation in feature space builds an information potential field flagging OOD samples on CIFAR-10 vs SVHN.Mon, 04 Aug 2025 00:00:00 GMTMachine LearningUncertainty QuantificationOOD DetectionarXivAIhttps://mllog.dev/en/posts/deep-learning-clinical-trial/https://mllog.dev/en/posts/deep-learning-clinical-trial/A multimodal Clinical Longformer model predicts trial enrollment at R² 0.76 and forecasts recruitment duration with 7.52-month MAE versus 10.55 for fit-and-filter.Sat, 02 Aug 2025 00:00:00 GMTclinical trialsmachine learningdeep learninguncertaintyhttps://mllog.dev/en/posts/coda-consensus-driven-active-model-selection/https://mllog.dev/en/posts/coda-consensus-driven-active-model-selection/CODA picks the best model from many candidates via their agreement: Bayesian Dawid-Skene plus information gain wins 18 of 26 benchmarks with up to 70% fewer labels.Fri, 01 Aug 2025 00:00:00 GMTmachine learningactive learningmodel selectionBayesian inferencehttps://mllog.dev/en/posts/rlvmr-wzmocnione-uczenie-z-weryfikowalnymi-nagradzajacymi-meta-rozumowaniem/https://mllog.dev/en/posts/rlvmr-wzmocnione-uczenie-z-weryfikowalnymi-nagradzajacymi-meta-rozumowaniem/RLVMR adds a verifiable meta-reasoning reward to the environment reward, scoring the agent's reasoning trace so long-horizon agents take fewer semantic shortcuts.Thu, 31 Jul 2025 00:00:00 GMTAIadaptacyjna sensorykarlehttps://mllog.dev/en/posts/honey-origin-classification-review-en/https://mllog.dev/en/posts/honey-origin-classification-review-en/Random Forests read the 12-mineral fingerprint of 429 honey samples to name the source flower with 99.50% accuracy and the region of origin with 98.01%.Wed, 30 Jul 2025 00:00:00 GMThoneymachine-learningfood-authenticitymineral-profileshttps://mllog.dev/en/posts/call-center-ppo-vs-vi-en/https://mllog.dev/en/posts/call-center-ppo-vs-vi-en/PPO beats classical Value Iteration at call routing in a discrete-event call centre simulation: shortest customer wait, least agent idle time, highest RL reward.Sat, 26 Jul 2025 00:00:00 GMTReinforcement LearningPPOValue IterationCall CenterArtificial Intelligencehttps://mllog.dev/en/posts/efficient-geometrically-smart-linear-memory-se2-invariant-attention/https://mllog.dev/en/posts/efficient-geometrically-smart-linear-memory-se2-invariant-attention/Fourier-encoded relative pose lets SE(2)-invariant attention use linear rather than quadratic memory, with error below 0.001. Best paper at the RSS 2025 workshop.Fri, 25 Jul 2025 00:00:00 GMTMachine LearningTransformersGeometric AIPopular Sciencehttps://mllog.dev/en/posts/ai-engine-skin-cancer/https://mllog.dev/en/posts/ai-engine-skin-cancer/A MobileNetV2 skin-lesion classifier compressed with TensorRT runs on a Jetson Orin Nano at 41% of its original size, scoring 87.2% F1 with 93% lower power in INT8.Thu, 24 Jul 2025 00:00:00 GMTAIMobileNetV2skin cancerwearable deviceshttps://mllog.dev/en/posts/sophia-enhancing-slow-thinking/https://mllog.dev/en/posts/sophia-enhancing-slow-thinking/SOPHIA mixes on-policy visual grounding with off-policy reasoning chains and backward reward propagation, adding 8.5 points of average accuracy to InternVL3.0-38B.Wed, 23 Jul 2025 00:00:00 GMTAILVLMreinforcement learningSOPHIAhttps://mllog.dev/en/posts/the-role-of-ai-in-managing-satellite-constellations/https://mllog.dev/en/posts/the-role-of-ai-in-managing-satellite-constellations/ConstellAI, an ESA project, uses reinforcement learning for routing in satellite mega-constellations, beating Dijkstra on delay and predicting bandwidth demand.Tue, 22 Jul 2025 00:00:00 GMTConstellAIReinforcement LearningSatelliteshttps://mllog.dev/en/posts/fundamental-limitations-dual-cvar-decompositions/https://mllog.dev/en/posts/fundamental-limitations-dual-cvar-decompositions/Godbout and Durand show the two dual CVaR formulas can share no feasible risk assignment, and build an MDP where no policy is optimal at every risk level.Mon, 21 Jul 2025 00:00:00 GMTMarkov Decision ProcessesCVaRdual decompositionhttps://mllog.dev/en/posts/pinfm-foundation-model-user-activity/https://mllog.dev/en/posts/pinfm-foundation-model-user-activity/PinFM, Pinterest's 20B-parameter sequence transformer, uses deduplicated cross-attention and int4 embeddings to lift throughput 600% and engagement 20%.Sun, 20 Jul 2025 00:00:00 GMThttps://mllog.dev/en/posts/gradnetot-method/https://mllog.dev/en/posts/gradnetot-method/GradNetOT learns optimal transport maps as the gradient of a convex potential, using a monotone gradient network trained on the Monge–Ampère residual.Sat, 19 Jul 2025 00:00:00 GMTGradNetOTOptimal TransportNeural Networkshttps://mllog.dev/en/posts/learning-on-the-border-ntk/https://mllog.dev/en/posts/learning-on-the-border-ntk/A larger learning rate leaves the NTK better aligned with the labels at the Edge of Stability: Kernel Target Alignment jumps exactly when sharpness drops.Fri, 18 Jul 2025 00:00:00 GMTpopularizationmachine learningNTKEdge of Stabilityhttps://mllog.dev/en/posts/riemannlora-unified-riemannian-framework/https://mllog.dev/en/posts/riemannlora-unified-riemannian-framework/RiemannLoRA places LoRA updates on the manifold of rank-r matrices, erasing factorization ambiguity: half the fine-tuning steps and up to 1.2 BLEU more.Thu, 17 Jul 2025 00:00:00 GMTMachine LearningRiemannian OptimizationLoRAFine-tuninghttps://mllog.dev/en/posts/nn-model-of-complementary-learning-systems/https://mllog.dev/en/posts/nn-model-of-complementary-learning-systems/A complementary learning systems model pairs a VAE with a modern Hopfield network for pattern separation and completion, hitting 90% average accuracy on Split-MNIST.Wed, 16 Jul 2025 00:00:00 GMThttps://mllog.dev/en/posts/target-polish/https://mllog.dev/en/posts/target-polish/Target Polish cleans data before factorizing it: clip X into a delta band around the current WH estimate, refactorize, repeat. Robust NMF for matrices and tensors.Tue, 15 Jul 2025 00:00:00 GMTNMFartificial intelligencemachine learningdata analysistarget polishhttps://mllog.dev/en/posts/optimistic-exploration-risk-averse-crl/https://mllog.dev/en/posts/optimistic-exploration-risk-averse-crl/ORAC maximizes an upper confidence bound on reward while minimizing a lower bound on cost, with an adaptive penalty, tested on Safety-Gymnasium and CityLearn.Mon, 14 Jul 2025 00:00:00 GMThttps://mllog.dev/en/posts/adaptive-sensing/https://mllog.dev/en/posts/adaptive-sensing/Adaptive sensing beats brute-force scale: EfficientNet-B0 paired with sensors that tune themselves to the scene outperforms OpenCLIP-H on perception tasks.Sun, 13 Jul 2025 00:00:00 GMTAIadaptive sensingmachine learninghttps://mllog.dev/en/posts/hgmp-revolution-graph-prompt-learning/https://mllog.dev/en/posts/hgmp-revolution-graph-prompt-learning/HGMP recasts node, edge and ranking tasks as graph-level prompts, combining contrastive pre-training with per-type feature prompts for heterogeneous graph learning.Sat, 12 Jul 2025 00:00:00 GMTgraph learningprompt learningmulti-taskAIresearchhttps://mllog.dev/en/posts/apexoracle-forecast-antibiotics/https://mllog.dev/en/posts/apexoracle-forecast-antibiotics/ApexOracle fuses Evo2 genome embeddings with a SELFIES diffusion model to predict MIC and design new antibiotics, with R2 gains up to 15% on P. aeruginosa.Fri, 11 Jul 2025 00:00:00 GMTMLAntibioticsDrug DiscoveryDiffusion Modelshttps://mllog.dev/en/posts/helo-multimodal-emotion-recognition/https://mllog.dev/en/posts/helo-multimodal-emotion-recognition/HeLo predicts a distribution over emotions instead of one label, fusing ECG and video via cross-attention, entropic optimal transport and a label-correlation loss.Thu, 10 Jul 2025 00:00:00 GMTmultimodalemotionmachine learninghttps://mllog.dev/en/posts/modern-methods-in-associative-memory/https://mllog.dev/en/posts/modern-methods-in-associative-memory/Hopfield networks store only 0.14N patterns; dense variants reach polynomial capacity and modern ones exponential, with a Lagrangian view linking them to attention.Wed, 09 Jul 2025 00:00:00 GMTAssociative MemoryHopfield NetworksTransformersDiffusion Modelshttps://mllog.dev/en/posts/quest-robust-quantile-estimation/https://mllog.dev/en/posts/quest-robust-quantile-estimation/QuEst blends a small gold sample with large model-generated imputations to estimate quantiles and CVaR, canceling simulator bias and minimizing sampling variance.Tue, 08 Jul 2025 00:00:00 GMTQuEstQuantile EstimationStatisticsMachine Learninghttps://mllog.dev/en/posts/retrysql-self-correction-in-sql-generation/https://mllog.dev/en/posts/retrysql-self-correction-in-sql-generation/RetrySQL injects deliberate wrong reasoning steps and a BACK token into text-to-SQL pretraining, lifting OpenCoder 1.5B execution accuracy by 4 points toward GPT-4o.Mon, 07 Jul 2025 00:00:00 GMTRetrySQLtext-to-SQLLLMmachine-learninghttps://mllog.dev/en/posts/mvho-ib-brain-disorder-diagnosis/https://mllog.dev/en/posts/mvho-ib-brain-disorder-diagnosis/MvHo-IB diagnoses brain disorders from rs-fMRI by adding triplet O-information to pairwise connectivity, gaining up to 7% accuracy on ABIDE, ADHD-200 and COBRE.Sun, 06 Jul 2025 00:00:00 GMTfMRIAIdeep learninginformation theoryneuroimaginghttps://mllog.dev/en/posts/multi-level-step-by-stepguidelines/https://mllog.dev/en/posts/multi-level-step-by-stepguidelines/StepHint splits reasoning chains by end-of-step token probability, then feeds RL agents hints at several levels, easing near-miss errors and stalled exploration.Sat, 05 Jul 2025 00:00:00 GMThttps://mllog.dev/en/posts/blog_xgboost/https://mllog.dev/en/posts/blog_xgboost/XGBoost on 341 temporal, spatial and network features predicts e-scooter demand 27-49% better than rivals and up to 90% better than Prophet, node centrality on top.Fri, 04 Jul 2025 00:00:00 GMTXGBoostAImachine learninghttps://mllog.dev/en/posts/xgboost-how-to-predict-demand/https://mllog.dev/en/posts/xgboost-how-to-predict-demand/Ghost nodes add fake output classes: softmax runs over 13 outputs while loss counts only the 10 real ones, giving gradients room to escape flat regions early.Thu, 03 Jul 2025 00:00:00 GMTghostAImachine learningtreeshttps://mllog.dev/en/posts/math_datasheets/https://mllog.dev/en/posts/math_datasheets/A neural-mechanistic hybrid forecasts epidemics from differentially private mobility and spending data, still beating public-health-only models under strong noise.Tue, 01 Jul 2025 00:00:00 GMTprywatnośćAInauczanie maszynoweepidemiehttps://mllog.dev/en/posts/blog_armor/https://mllog.dev/en/posts/blog_armor/ARMOR trains a teacher VAE on privileged attack data and an LSTM student on raw sensor history, giving UAVs zero-shot robustness to unseen GPS spoofing and jamming.Mon, 30 Jun 2025 00:00:00 GMThttps://mllog.dev/en/posts/mind2web-agent-based-web-search/https://mllog.dev/en/posts/mind2web-agent-based-web-search/Mind2Web 2 grades agentic web search on 130 tasks with an Agent-as-a-Judge rubric; the best of nine systems reaches 50-70% of human quality at twice the speed.Sun, 29 Jun 2025 00:00:00 GMTAIsearchingmind2webhttps://mllog.dev/en/posts/hfex-machine-that-discovers-the-laws-of-physics/https://mllog.dev/en/posts/hfex-machine-that-discovers-the-laws-of-physics/H-FEX searches symbolic expressions for a system's Hamiltonian, recovering exp(-1.0003p² - 1.1002q⁴) almost exactly and beating SINDy on long-term energy drift.Sat, 28 Jun 2025 00:00:00 GMTphysicsAIsymbolic learningH-FEXhttps://mllog.dev/en/posts/bandit_learning_artykul/https://mllog.dev/en/posts/bandit_learning_artykul/Bandit learning has no VC-dimension analogue: a class can need only 2 queries yet stay intractable unless RP = NP. With a Rust epsilon-greedy A/B/n example.Fri, 27 Jun 2025 00:00:00 GMT